The AI bottleneck isn't chips. Two ex-miners are cashing in.

For two years the story of the AI buildout has been about one shortage: GPUs. But the constraint that now decides who builds first has quietly moved upstream — to electricity, substations, and the right to plug into the grid. And the companies sitting on that scarce asset are not data-center developers. They are former Bitcoin miners.

Why the AI Bottleneck Is Now Power, Not Chips

The binding constraint in AI infrastructure has shifted from chip availability to "time-to-power" and "time-to-compute" — and a subset of public Bitcoin miners already controls the scarce input. That is the core argument Bernstein analyst Gautam Chhugani made when the firm initiated equity coverage on TeraWulf (Nasdaq: WULF) and Cipher Digital (Nasdaq: CIFR) in early June 2026, rating both Outperform with a $36 target on TeraWulf and a $32 target on Cipher . The thesis reframes these firms entirely: not a hash-rate story, but a power-and-real-estate one.

Quick Answer: The AI bottleneck has moved from GPUs to electricity. Bernstein argues former Bitcoin miners already own the scarce asset — powered land — and the covered cohort has contracted roughly 6 GW to hyperscalers across 17 deals worth $110B+ over two years, about 10% of U.S. AI data centers under construction.

The logic rests on what miners physically hold. Brownfield mining campuses come with energized substations, legacy interconnection queue positions, operating crews, and multi-gigawatt power pipelines — the exact stack a greenfield developer would otherwise need an estimated four to seven years to assemble from scratch . Because miners spent years securing cheap, large blocks of electricity, that historical edge maps directly onto today's scarcity. They can lease AI-ready capacity to hyperscalers and "neocloud" operators faster than anyone building on raw land.

The scale is already material. Bernstein notes the covered miner cohort has contracted roughly 6 gigawatts of power to hyperscalers and neoclouds across 17 deals worth more than $110 billion over two years — equivalent to about 10% of all U.S. AI data centers currently under construction — with a planned aggregate portfolio approaching 30 GW . Named end customers across the space include Google, Amazon, Microsoft, Nvidia and CoreWeave.

"The power landlords of AI," — Gautam Chhugani, analyst at Bernstein, describing miners that can monetize energized grid capacity faster than greenfield developers (source: Decrypt).

These remain sell-side forecasts rather than company guidance, and the original Bernstein note is not a public primary source. But the call did not arrive in isolation: it follows bullish initiations from Morgan Stanley (Overweight, February 2026) and Jefferies (Buy, May 2026) . The sections that follow unpack what that powered land is actually worth.

The 9x Forecast Unpacked: $1.2B to $10.7B in AI Revenue by 2030

The headline number is a roughly ninefold jump in four years. Bernstein projects aggregate AI revenue across its covered miner universe rising from about $1.2 billion in 2026 to about $10.7 billion by 2030 . That trajectory rests on contracted, long-duration lease income rather than mining margins — and the analyst frames it as the value of powered land finally being priced in, not a hash-rate rerating .

The growth is concentrated in the two names Bernstein initiated. By 2030 it models TeraWulf reaching about $1.7 billion in AI revenue — one report cites a ~163% CAGR off a roughly $14 million 2025 base — at site-NOI/EBITDA margins near 84–85% . Cipher Digital is modeled at about $1.2 billion by 2030 at margins near 93%, the higher figure reflecting its more hyperscaler-weighted client base . Named end customers across the cohort include Google, Amazon, Microsoft, Nvidia and CoreWeave .

The financial backbone matters as much as the topline. Bernstein describes the revenue as long-duration, credit-enhanced, take-or-pay colocation leases — tenants pay whether or not they draw the capacity — with project financing said to cover roughly 75–85% of construction costs . Across the broader cohort, the analyst counts about 6 gigawatts already contracted to hyperscalers and neoclouds — 17 deals worth more than $110 billion over two years, equivalent to roughly 10% of U.S. AI data centers currently under construction — against a planned aggregate power portfolio of about 30 GW .

The call also sits within a cluster of bullish sell-side initiations rather than standing alone. Morgan Stanley moved to Overweight in February 2026 with targets in the $37–$38 range, and Jefferies issued Buy ratings in May 2026 at $32 on Cipher and $28 on TeraWulf . The convergence is notable, but every figure here is a sell-side projection keyed to delivery dates that mostly fall after 2026 — so the $10.7 billion endpoint is a forecast of contracted potential, not booked revenue.

TeraWulf (WULF): From Hash Rate to 1+ GW of Contracted AI Real Estate

TeraWulf is the clearest example of a Bitcoin miner re-engineering itself into a contracted AI landlord, and its income statement is now mid-pivot. In its FY 2025 results, the company reported just $16.9 million of HPC lease revenue against $151.6 million of digital-asset revenue . One quarter later the mix had inverted at the margin: Q1 2026 revenue of $34.0 million included $21.0 million of HPC lease revenue — roughly 60% of the total — alongside 60 MW of operational critical IT capacity serving Core42 . That is the inflection Bernstein's $36 target is built on, even though most contracted megawatts are not yet generating rent.

Quick Answer: TeraWulf is converting cheap mining power into long-duration AI hosting leases. As of February 2026 it held 522 critical IT MW under contract, representing more than $12.8 billion of contracted revenue and $6.5 billion of completed long-term financings, with about $3.1 billion of cash and restricted cash at March 31, 2026 .

The anchor of that platform is the Fluidstack agreement signed in mid-August 2025 at the Lake Mariner site in Western New York. The deal began at 200+ MW (roughly 250 MW gross), ran 10 years, and was valued at about $3.7 billion initially — up to roughly $8.7 billion with two optional five-year extensions — at an expected build cost of $8–10 million per MW and site-NOI margins near 85% . A subsequent 160 MW CB-5 expansion lifted contracted capacity further .

What makes the structure unusual is who stands behind it. Google does not lease the capacity directly; instead it backstops $1.8 billion of Fluidstack's obligations and holds warrants for about 41 million shares, equal to roughly 8% on a pro-forma basis . That credit enhancement is the linchpin of the bull case — it lets a sub-investment-grade tenant carry hyperscaler-grade risk, which in turn unlocks project financing. As Bernstein framed the broader cohort in its initiation note, these miners have become "the power landlords of AI," monetizing energized sites rather than hash rate .

Beyond Lake Mariner, TeraWulf is widening the platform on two fronts:

• Abernathy HPC Campus (Texas): a majority-owned Fluidstack joint venture designed for 168 critical IT MW under a 25-year lease with annual escalators and Google support .
• Muskie Data Campus (Eastern Kentucky): acquired May 26, 2026, expected to support more than 1 GW over time — an initial 500 MW ramp from the second half of 2028 and another 500 MW from the second half of 2030 — anchored by a 345 kV substation tied to an existing 765 kV transmission network .

The Kentucky acquisition is telling: the value sits less in the land than in the interconnection rights, because tying into an energized 765 kV network is precisely the time-to-power advantage greenfield developers cannot replicate quickly. The financial caveat remains stark, though. The 2025 10-K still showed a $661.4 million net loss against $168.5 million of total revenue , a reminder that the contracted backlog is potential cash flow, not realized earnings — and that the gap closes only as 2026–2028 delivery windows are met.

Cipher Digital (CIFR): Hyperscaler Concentration and the 93% Margin Case

Cipher Digital is the more hyperscaler-concentrated of Bernstein's two "power landlords," and that concentration is the entire margin story. Bernstein projects Cipher reaching roughly $1.2 billion in AI revenue by 2030 at site-NOI margins near 93% — versus about 85% for TeraWulf — because Cipher contracts more directly with end hyperscalers rather than routing capacity through a neocloud intermediary . The clearest signal of that pivot was the rebrand from Cipher Mining to Cipher Digital on February 20, 2026, made explicitly to flag a strategic shift toward HPC/AI development and stable, long-duration lease cash flows .

As of its February 24, 2026 update, Cipher reported 600 MW gross of contracted HPC capacity split across two leases: a 15-year, 300 MW agreement with Amazon Web Services and a 10-year, 300 MW Fluidstack/Google lease, alongside three completed bond offerings totaling $3.73 billion of proceeds to fund construction . The AWS lease is the differentiator: approximately $5.5 billion over 15 years for 300 MW, with delivery in 2026 and rent expected to begin in August 2026 . Contracting directly with an investment-grade counterparty like AWS — rather than a startup neocloud — is what underpins the premium margin assumption.

The Barber Lake deal follows the now-familiar credit-enhancement template. It began September 25, 2025 at 168 critical IT MW, worth roughly $3.0 billion over 10 years, with Google backstopping $1.4 billion of Fluidstack's obligations and receiving warrants for about 24 million shares, near 5.4% pro forma . A November 20, 2025 expansion added 56 MW gross — about 39 critical IT MW — lifting the partnership's initial 10-year contracted revenue to roughly $3.8 billion and Google's backstop to about $1.73 billion . SoftBank Group had already signaled conviction earlier, investing $50 million in January 2025 to support the HPC build-out .

The underlying asset base is broader than the two headline leases suggest. Cipher's 2025 10-K describes a 4.2 GW portfolio across 10 sites, 600 MW of HPC under development for hyperscalers, about 207 MW still operating at one Texas Bitcoin mining site, and a 3.4 GW pipeline beyond what is already contracted . The clearest near-term expansion candidate is the 1 GW West Texas Colchis site, which holds an AEP interconnection and targets a 2028 in-service date .

The financial reality, as with TeraWulf, is that the backlog is still mostly contractual rather than realized. In Q1 2026 Cipher posted just $35 million of revenue and adjusted EBITDA of negative $48 million, even as it signed a third 15-year campus lease with an undisclosed investment-grade hyperscale tenant . That third lease extends the same playbook — long-duration, top-tier counterparty — but the 93% margin thesis only converts to cash once the AWS rent commences in August 2026 and the 2026–2028 delivery windows are actually met.

Deal Architecture: How Take-or-Pay Leases and Credit Enhancement Actually Work

The repeatable engine behind these forecasts is a take-or-pay colocation lease: a long-duration contract, typically 10 to 25 years, under which a hyperscaler or neocloud tenant pays fixed rent on contracted capacity whether or not it fully uses it, with annual escalators that lift rent over the term. Project financing is structured to cover roughly 75–85% of construction cost , so contracted cash flows largely service the construction debt while equity retains the upside from NOI margin and escalation. That financial shape — predictable rent against amortizing debt — is what lets analysts treat these as real-estate cash flows rather than commodity mining revenue.

Credit enhancement is what makes the structure bankable. Neocloud tenants such as Fluidstack do not carry investment-grade balance sheets, so Google steps in to backstop the tenant's payment obligations in exchange for warrants on the developer's equity — adding top-tier credit quality without appearing as the direct lease counterparty. At TeraWulf's Lake Mariner, Google backstopped $1.8 billion of Fluidstack's obligations and took warrants for about 41 million shares (~8% pro forma) . At Cipher's Barber Lake, Google backstopped $1.4 billion and received warrants for roughly 24 million shares (~5.4% pro forma) .

A measurement distinction is essential when comparing these deals across press releases and SEC filings. Gross MW is nameplate electrical capacity; critical IT MW is the power actually delivered to compute, typically about 65–70% of gross after cooling and overhead. The gap explains why the same campus can be quoted at 600 MW gross or roughly 168–300 critical IT MW depending on the document, so cross-source comparisons require checking which metric is cited and on what date.

Sources: TeraWulf, 2025-08 · Cipher/Fluidstack expansion, 2025-11

The build economics scale cleanly once that structure is in place. Total project cost runs about $8–10 million per MW , and at site-NOI margins of 84–93% reported across the two operators, a fully ramped 300 MW campus throws off roughly $240–280 million in annual site NOI . The arithmetic is straightforward; the execution — energizing capacity on schedule and meeting hyperscaler reliability standards — is where the cash flows are actually won or lost.

The Macro Backdrop: Why U.S. Powered Land Became Scarce Overnight

The scarcity those margins depend on is structural, not cyclical. The bottleneck for AI buildout has moved from silicon to electricity: large, contiguous blocks of energized, grid-connected power are now the binding constraint, and the United States is short of them. The International Energy Agency's Energy and AI analysis finds that the U.S. and China together account for nearly 80% of global data-center electricity-consumption growth through 2030, with U.S. data centers alone making up close to half of all U.S. electricity-demand growth over the same period [13]. Demand is concentrating into exactly the resource that takes longest to add.

Three frictions explain why powered land cannot simply be built to order:

• Interconnection backlogs. Queues to connect new load and generation to the grid now stretch beyond five years in several key U.S. markets. A miner's legacy queue position and energized substation rights were accumulated over years of operation — they are not replicable on a hyperscaler's timeline, which is the structural moat behind the "power landlord" framing.
• Time-to-power asymmetry. Greenfield data-center development typically runs four to seven years from site selection to energized capacity. Converting an existing miner campus with live interconnection, substations, and operating crews compresses that to roughly 18–24 months — the difference between delivering capacity into the 2026–2027 demand window and missing it entirely.
• Hyperscaler capex. Google, AWS, Microsoft, and Meta have outlined combined 2025–2026 capital plans exceeding $300 billion, weighted heavily toward data-center buildout. The same names — Google, Amazon, Microsoft, Nvidia, and CoreWeave — appear as end customers across the contracted-miner cohort [1], demand that is outpacing the available supply of powered land.

This is the asymmetry Bernstein leans on in its coverage. As analyst Gautam Chhugani frames the call, the miners that pivoted early are "the power landlords of AI" — owners of scarce, energized real estate rather than commodity hash rate [2]. The cohort has already contracted out roughly 6 GW across 17 deals worth more than $110 billion in two years, about 10% of U.S. AI data centers currently under construction [1].

The takeaway for a trader: the value here is the energized substation and the queue position, not the ASICs. As long as interconnection timelines stay measured in years while hyperscaler capex is measured in hundreds of billions, the spread between time-to-power and time-to-demand is what these operators are monetizing.

Execution Risks: What Has to Go Right for the 9x Call to Hold

The 9x revenue call rests on capacity that is mostly still on paper, not on the grid. The single biggest execution risk is timing: the bulk of contracted megawatts across TeraWulf and Cipher Digital are not yet operational, with delivery windows clustered in 2026–2027 and stretching well beyond. As of Q1 2026, TeraWulf had only 60 MW of operational critical IT capacity serving Core42 , against the 522 critical IT MW it has contracted . Every gigawatt of Bernstein's $10.7 billion 2030 figure depends on construction finishing on schedule.

Both companies are GAAP-loss businesses funding that buildout with leverage that contracted revenue must later service. TeraWulf reported a $661.4 million net loss for FY 2025 , and Cipher posted adjusted EBITDA of negative $48 million on $35 million of quarterly revenue in Q1 2026 . With project financing covering roughly 75–85% of construction costs , any gap between cash burn and lease commencement pressures balance-sheet reserves.

The risks worth tracking before the EBITDA inflection:

• Permitting and grid upgrades. Substation and transmission work can slip 12–24 months. The later phases of TeraWulf's Muskie campus — a second 500 MW from 2H 2030 — and Cipher's 1 GW West Texas Colchis site, which targets 2028 on AEP interconnection , sit on utility timelines outside management's control.
• Hyperscaler reliability standards. Carrier-grade uptime at scale is operationally distinct from Bitcoin mining, where downtime is merely lost hash. Cipher's AWS lease, with rent expected to begin August 2026 , is the first major proof point for whether converted miner campuses can sustain tier-3/4 service levels.
• Demand reversal. Because the thesis is explicitly decoupled from Bitcoin price, an AI capex contraction or hyperscaler consolidation would directly shrink future contracted capacity and hit asset valuations .
• Lease commencement timing. Forecasts are highly sensitive to go-live dates. Slippage in the AWS start or in TeraWulf's 160 MW CB-5 expansion would push back the revenue ramp and the EBITDA turn.

For a trader, the practical read is that contracts and credit backstops de-risk the demand side, but not the build side. The proof points to watch through 2026–2027 are operational MW coming online on schedule, the AWS commencement clearing on time, and financing costs staying contained — not new headline deals. Until contracted megawatts convert to energized, rent-paying capacity, the 9x figure remains a sell-side projection rather than realized cash flow.

2030 Outlook: Bull and Bear Cases Side by Side

The 2030 outlook splits cleanly along a single fault line: whether contracted megawatts convert to energized, rent-paying capacity on schedule. Sell-side coverage is unusually aligned for two thinly-followed names. Bernstein rates both Outperform with a $36 target on TeraWulf and $32 on Cipher Digital , Morgan Stanley went Overweight with $37–$38 targets in February 2026, and Jefferies issued Buy ratings in May 2026 at $32 on Cipher and $28 on TeraWulf . That consensus is constructive, but it rests on forecasts, not realized cash flow.

Quick Answer: The bull case sees contracted capacity deliver on time, lifting TeraWulf to ~$1.7B AI revenue at ~85% site-NOI margins and Cipher to ~$1.2B at ~93%, with the covered cohort reaching ~$10.7B by 2030. The bear case is delay: slipped buildouts, a cooling AI capex cycle, and pricier financing.

The aggregate prize is large: Bernstein projects covered-miner AI revenue rising roughly ninefold from about $1.2 billion in 2026 to $10.7 billion by 2030 . The bear scenario does not require demand to collapse — only for the AI capex cycle to moderate after its 2026–2027 peak, for financing costs to rise into the buildout, or for hyperscalers to favor in-house capacity over leased megawatts.

For traders, the near-term watch list is concrete rather than speculative. Four datable catalysts will signal whether the thesis is tracking: Cipher's AWS rent commencement, expected to begin in August 2026 ; TeraWulf's 160 MW CB-5 expansion go-live ; the permitting timeline for the Muskie campus in Eastern Kentucky, slated for an initial 500 MW ramp from 2H 2028 ; and whether either firm signs a fourth major hyperscaler lease before year-end. The takeaway: treat the $10.7 billion figure as a milestone to verify, not a destination to assume. Energized, rent-paying megawatts — not new headline deals — are the metric that separates the bull case from the bear case.

Frequently asked questions

What does 'power landlord' mean in the context of Bitcoin miners?

A "power landlord" is a former Bitcoin miner that now earns revenue by leasing energized land and infrastructure — not hash rate — to AI tenants under long-term contracts. The scarce asset is brownfield power: grid interconnection rights, energized substations, and operating crews that take years to permit and build. Miners accumulated these assets before AI demand surged, giving them a head start over greenfield data-center developers. Bernstein coined the term in early June 2026 when it initiated coverage on TeraWulf and Cipher Digital, noting the covered cohort had contracted roughly 6 GW to hyperscalers and neoclouds across 17 deals worth more than $110 billion .

What is Bernstein's price target for TeraWulf (WULF) and Cipher Digital (CIFR)?

Bernstein set a $36 price target on TeraWulf and a $32 target on Cipher Digital, rating both Outperform when it initiated coverage in June 2026 . It was not alone: Morgan Stanley initiated Overweight in February 2026 with $37–$38 targets, and Jefferies issued Buy-equivalent ratings in May 2026 at $32 for Cipher and $28 for TeraWulf . Important caveat: these are sell-side analyst forecasts, not company guidance, and price targets reflect analyst modeling assumptions rather than committed financial outcomes.

Why does Cipher Digital have higher projected margins than TeraWulf?

Cipher Digital's client mix is more directly hyperscaler-weighted, which supports Bernstein's projected ~93% site-NOI/EBITDA margin versus ~85% for TeraWulf . Cipher's roughly $5.5 billion, 15-year lease for 300 MW is a direct contract with Amazon Web Services . TeraWulf's largest agreements, by contrast, run through Fluidstack — a neocloud intermediary — which adds a counterparty layer between the landlord and the end demand. Direct hyperscaler leases typically command tighter credit spreads and more predictable payment terms, which is why an investment-grade tenant like AWS underpins a thinner-cost, higher-margin profile in the model.

How does Google's credit enhancement work in these AI lease deals?

Google credit-enhances these leases by backstopping the neocloud tenant's payment obligations — agreeing to cover Fluidstack's lease payments if it defaults — in exchange for warrants on the developer's equity. This injects investment-grade credit into the deal without Google signing the lease itself. The backstop scales with contracted capacity: Google provided a $1.8 billion backstop plus warrants for roughly 41 million TeraWulf shares (~8% pro forma) at Lake Mariner , and a $1.4 billion backstop — later lifted to about $1.73 billion — plus warrants for roughly 24 million Cipher shares (~5.4% pro forma) at Barber Lake .

When will TeraWulf and Cipher Digital become EBITDA-positive?

Both firms remain EBITDA-negative as of the first quarter of 2026 and depend on contracted megawatts going live to inflect. Cipher reported adjusted EBITDA of negative $48 million on $35 million of revenue in Q1 2026, while still in its buildout phase . TeraWulf's transition was further along on the revenue line — HPC lease income reached $21.0 million of its $34.0 million Q1 2026 total, about 60% . Inflection hinges on the 2026–2027 delivery tranche ramping: Cipher's AWS rent was expected to begin in August 2026 . Lease-commencement timing is the key variable, so treat any positive-EBITDA date as conditional on energized, rent-paying capacity.